Psychosocial interventions for Alzheimer’s disease cognitive symptoms: a Bayesian network meta-analysis

For the two network meta-analyses, we searched MEDLINE, EMBASE and PsycINFO through the OVID database, and searched CENTRAL through the Cochrane Library. We searched studies published from inception to August 31, 2017, and compared any psychosocial interventions for cognitive symptoms in patients with AD (Additional file 1).

Inclusion criteria and exclusion criteria are summarized in Table 1.

We included the following psychosocial interventions with usual care as the control: home-based exercise (HE), group exercise (GE), walking programs (WP), reminiscence therapy (RT), and art therapy (AT). We also included the combination of psychosocial interventions and ChEIs with ChEIs as a positive control: the combination of cognitive stimulation treatment and acetylcholinesterase inhibitor (CST + ChEIs), the combination of mindfulness-based Alzheimer’s stimulation and acetylcholinesterase inhibitor (MBAS+ChEIs), the combination of progressive muscle relaxation and acetylcholinesterase inhibitor (PMR + ChEIs), and the combination of cognitive training and acetylcholinesterase inhibitor (CT + ChEIs).

Three investigators (YTD, LLM, JXC) independently selected the studies. The review of the main reports and supplementary materials, the extractions of the relevant information from the included trials with a predetermined data extraction sheet, and the assessments of the risk of bias with the Cochrane risk of bias tool were independently performed by three investigators (CXW, JLL, YZ). Any disagreements were resolved through discussion. When they did not reach a consensus, the final decision about each question was made by other investigators within the review team (JJW, CZT, PJR).

We evaluated the quality of the included studies with the Cochrane Collaboration Recommendations assessment tool. The tool for assessing 7 domains, including random sequence generation, allocation concealment, blinding of participants and personnel, blinding (or masking) of outcomes assessors, incomplete outcome data, selective reporting and other bias is described in the Cochrane Handbook for Systematic Reviews of Interventions (see details at http://​training.​cochrane.​org/​handbook). Based on these items, studies in which the key domains were all low-risk were considered low-risk, while the remainder were deemed high-risk or unclear-risk, depending on the number of key domains of high or unclear risk.

A network meta-analysis with a Bayesian framework with Aggregate Data Drug Information System (ADDIS, version 1.16.8) was conducted to assess the cognitive outcomes of psychosocial interventions. This software is based on the Bayesian framework and the Markov chain Monte Carlo method which can evaluate a priori and process research data. We used a random-effects model to analyze the effect sizes in this study. The effect sizes for continuous outcomes were the mean difference (MD). Consistency and inconsistency were the two models used to estimate the effect size in ADDIS. A consistency assessment drew conclusions on effect sizes of the included interventions and estimated the ranking probabilities for all the interventions. The consistency test was judged by node-splitting analysis and an inconsistence model. When the p-value of the node-splitting analysis was greater than 0.05, a consistency mode was selected [11]. Otherwise, an inconsistency model was used. Potential scale reduction factor (PSRF) was used to evaluate the convergence of the model. The closer the PSRF value was to 1, the better the convergence. The convergence of the model was still acceptable if the PSRF value was less than 1.2. For each intervention, we estimated the ranking probabilities for each treatment at each possible rank.

We ran pair-wise meta-analyses to compare the compliance of different psychosocial therapies because the data included in our study were insufficient for statistical analysis of the network meta-analysis. We conducted the pair-wise meta-analysis with the fixed-effects model with Review Manager (RevMan, v 5.3). The odds ratio (OR) was calculated for dichotomous outcomes (compliance), with 95% credible intervals (CI). We assessed statistical heterogeneity in the pair-wise comparison with an I² statistic and the p-value.

In total, 8445 citations published between 1981 and August 31, 2017 was identified by the search. After removing duplicates and unrelated articles, 10 articles describing 11 RCTs including 682 patients were eligible for further quantitative analyses. A flow chart of the specific screening procedures is shown in Fig. 1. The baseline characteristics of the studies were also extracted (Table 2).

We evaluated the quality of included studies with the Cochrane Collaboration Recommendations assessment tools [12]. Among 10 trials, 8 studies (80%) described a random component in the sequence generation process such as a computer-generated random number or a random number table. The others did not permit judgment of ‘low risk’ or ‘high risk’ due to insufficient information about the sequence generation process. Allocation concealment was performed using an appropriately sealed method in 70% (7) of the studies, while 30% (3) either did not describe concrete methods or used an inappropriate allocation concealment method. In performance bias, 70% (7) of the included trials reported the methods of blinding for both participants and personnel. In detection bias, 30% (3) of the outcome assessors in the studies either could not be blinded or were unclear. In attrition bias, 9 studies were deemed to have low-risk outcome data (i.e. reported dropout rates within the range of statistical estimations, provided detailed explanations of dropout rates or performed intention-to-treat analysis). 1 study did not provide adequate information to judge the risk of missing data. Other risks were unclear due to insufficient information in 5 studies. Overall, 1 study was considered high risk and 5 were considered low risk, while 4 were considered unclear risk. A detailed quality assessment is presented in Additional files 2 and 3.

A network meta-analysis was performed to compare and rank the included psychosocial interventions which used usual care as their control. The network of eligible comparisons for efficacy consisted of 6 studies and 5 treatments (2 arms of RT; 1 arm of AT, GE, HE, WP; 6 arms of control). The specific network is presented in Fig. 2a. The consistency model was selected for the subsequent network analyses. Meanwhile, the inconsistency model was used to test consistency. The results of the network meta-analysis for the primary outcomes are presented as a league table in Table 3. In terms of efficacy, WP was better than the control (SMD 4.98, 95% CI -0.07 to 10.00). This was the best psychosocial intervention for improving cognitive symptoms assessed by the Mini–Mental State Examination (MMSE). The results indicated that WP was significantly more effective than the other treatments in our study. The second and third most effective interventions were RT and HE. The ranking probability of treatments is presented in Fig. 2b and c

Consisting of 5 studies and 5 treatments (2 arms of CST + ChEIs; 1 arm of CT + ChEIs, MBAS+ChEIs, RMP + ChEIs; 5 arms of ChEIs), another network meta-analysis was run to assess the effectiveness of the combination of psychosocial interventions with ChEIs and simple ChEIs treatment. The specific network is presented in Fig. 3a. In this subsequent network analysis, we used a consistency model, while an inconsistency model was used to test consistency. The results of the network meta-analyses for the primary outcomes are presented as a league table in Table 4. Note that the curative effect, CT + ChEIs, the best combined intervention for improving cognitive symptoms assessed by MMSE, was better than the control (SMD 6.27, 95%CI -1.05 to 13.44). These results suggested that CT + ChEIs was significantly more effective than the other treatments. The second and third most effective interventions were CST + ChEIs and MBAS+ChEIs. The three most effective combined interventions were better than the ChEIs. The ranking probability of treatments is presented in Fig. 3b and c.

Additional file 4 shows the results of pair-wise meta-analyses of compliance for each intervention. The included studies which were not analyzed did not have missing data due to patient non-compliance. We can conclude that patients’ compliance with RT, WP and HE are better than GE and AT, and that patients’ compliance with CT + ChEIs, CST + ChEIs were better than those of MBAS+ChEIs, RMP + ChEIs and ChEIs.